Decision-Making on Robots using POMDPs and Answer Set Programming
Robots are increasingly becoming an integral part of many critical application sectors such as medicine, disaster rescue and the military. The deployment of mobile robots in real-world domains poses many complex challenges. Real-world domains are typically characterized by partial observability, non-deterministic action outcomes and unforeseen dynamic changes. Furthermore, it is not feasible for robots to process all sensory inputs using all available algorithms, making it a formidable challenge for robots to operate reliably and efficiently. Probabilistic planning using partially observable Markov decision processes (POMDPs) elegantly models the uncertainty in sensing and actuation on robots in the real-world. However, POMDP formulations of large complex domains are computationally intractable. In addition, POMDPs are not well-suited for knowledge representation or common sense reasoning. Answer set programming (ASP), on the other hand, is a non-monotonic logic programming paradigm that is appropriate for common sense reasoning, but cannot model the uncertainty in sensing and actuation. This research project investigates the use of a POMDP hierarchy in conjunction with ASP to enable mobile robots to reliably and efficiently localize target objects in indoor domains. Target objects are characterized by local image gradient features, and the existing POMDP hierarchy enables the robot to decide where to look, what images to process and how to process these images. The current project then focuses on robustly identifying the objects by observing them from different viewpoints. This is a complex decision-making challenge that is formulated as a two-layer POMDP. The bottom layer computes the sequence of viewpoints to be used to observe the target, while the top layer uses the information provided by the bottom layer to determine if the observations correspond to the desired target. The algorithms are implemented and evaluated in simulation and on wheeled robots in complex indoor domains.